Last week, Google put one of its engineers on administrative leave after he claimed to have encountered machine sentience on a dialogue agent named LaMDA. Because machine sentience is a staple of the movies, and because the dream of artificial personhood is as old as science itself, the story went viral, gathering far more attention than pretty much any story about natural-language processing (NLP) has ever received. That’s a shame. The notion that LaMDA is sentient is nonsense: LaMDA is no more conscious than a pocket calculator. More importantly, the silly fantasy of machine sentience has once again been allowed to dominate the artificial-intelligence conversation when much stranger and richer, and more potentially dangerous and beautiful, developments are under way.
The fact that LaMDA in particular has been the center of attention is, frankly, a little quaint. LaMDA is a dialogue agent. The purpose of dialogue agents is to convince you that you are talking with a person. Utterly convincing chatbots are far from groundbreaking tech at this point. Programs such as Project December are already capable of re-creating dead loved ones using NLP. But those simulations are no more alive than a photograph of your dead great-grandfather is.
Already, models exist that are more powerful and mystifying than LaMDA. LaMDA operates on up to 137 billion parameters, which are, speaking broadly, the patterns in language that a transformer-based NLP uses to create meaningful text prediction. Recently I spoke with the engineers who worked on Google’s latest language model, PaLM, which has 540 billion parameters and is capable of hundreds of separate tasks without being specifically trained to do them. It is a true artificial general intelligence, insofar as it can apply itself to different intellectual tasks without specific training “out of the box,” as it were.
Some of these tasks are obviously useful and potentially transformative. According to the engineers—and, to be clear, I did not see PaLM in action myself, because it is not a product—if you ask it a question in Bengali, it can answer in both Bengali and English. If you ask it to translate a piece of code from C to Python, it can do so. It can summarize text. It can explain jokes. Then there’s the function that has startled its own developers, and which requires a certain distance and intellectual coolness not to freak out over. PaLM can reason. Or, to be more precise—and precision very much matters here—PaLM can perform reason.
The method by which PaLM reasons is called “chain-of-thought prompting.” Sharan Narang, one of the engineers leading the development of PaLM, told me that large language models have never been very good at making logical leaps unless explicitly trained to do so. Giving a large language model the answer to a math problem and then asking it to replicate the means of solving that math problem tends not to work. But in chain-of-thought prompting, you explain the method of getting the answer instead of giving the answer itself. The approach is closer to teaching children than programming machines. “If you just told them the answer is 11, they would be confused. But if you broke it down, they do better,” Narang said.
Google illustrates the process in the following image:
Adding to the general weirdness of this property is the fact that Google’s engineers themselves do not understand how or why PaLM is capable of this function. The difference between PaLM and other models could be the brute computational power at play. It could be the fact that only 78 percent of the language PaLM was trained on is English, thus broadening the meanings available to PaLM as opposed to other large language models, such as GPT-3. Or it could be the fact that the engineers changed the way that they tokenize mathematical data in the inputs. The engineers have their guesses, but they themselves don’t feel that their guesses are better than anybody else’s. Put simply, PaLM “has demonstrated capabilities that we have not seen before,” Aakanksha Chowdhery, a member of the PaLM team who is as close as any engineer to understanding PaLM, told me.
None of this has anything to do with artificial consciousness, of course. “I don’t anthropomorphize,” Chowdhery said bluntly. “We are simply predicting language.” Artificial consciousness is a remote dream that remains firmly entrenched in science fiction, because we have no idea what human consciousness is; there is no functioning falsifiable thesis of consciousness, just a bunch of vague notions. And if there is no way to test for consciousness, there is no way to program it. You can ask an algorithm to do only what you tell it to do. All that we can come up with to compare machines with humans are little games, such as Turing’s imitation game, that ultimately prove nothing.
Where we’ve arrived instead is somewhere more foreign than artificial consciousness. In a strange way, a program like PaLM would be easier to comprehend if it simply were sentient. We at least know what the experience of consciousness entails. All of PaLM’s functions that I’ve described so far come from nothing more than text prediction. What word makes sense next? That’s it. That’s all. Why would that function result in such enormous leaps in the capacity to make meaning? This technology works by substrata that underlie not just all language but all meaning (or is there a difference?), and these substrata are fundamentally mysterious. PaLM may possess modalities that transcend our understanding. What does PaLM understand that we don’t know how to ask it?
